Professional Documents
Culture Documents
1.
Introduction
In the best of all morphosyntactically possible worlds, one might well suppose the
pieces of syntax and the pieces of morphology would be a one-to-one correspondence. In the real world, many factors impinge on this maximally simple relationship: allomorphy, accidental homophony, principled syncretism, non-cumulative
exponence, ... In this chapter, we concentrate on the first of these, aiming to
lay out the empirical boundaries that distinguish allomorphy from other sources
of variation and the theoretical issues that a fully fledged account of allomorphy
should, in our opinion, address.
What we mean by allomorphy is most simply expressed graphically. As a
preliminary, the exponence relation between a feature set and an exponent is represented as:
[F]
(1)
This is read as [F] has exponent ; for instance, if [F] is the feature specification
of the English present participle, then is /i/.
To serve breadth of readership, we understand feature set loosely, leaving it
to specific theoretical proposals, rather than to our definition, to determine whether
[F] is restricted to features bundles at terminal nodes or may include larger syntactic structures. Similarly, although we will use terminology familiar from postsyntactic theories of morphology, we regard such notation, and most of our discussion, as equally relevant to presyntactic (lexicalist) models. (For simplicity,
we will frequently adopt two notation devices: first, we use glossing labels rather
Acknowledgements to go here.
than actual features for [F], writing, e.g., [pl] rather than attempting to decide
between [plural], [+plural], [singular], etc.; and we use orthographic instead of
phonological strings for , writing e.g., ing for /i/.)
Given (1), we use allomorphy to refer to such exponence relations as:
1 Context1
Context
2
2
(2)
[F]
..
...
.
That is to say, [F] is said to exhibit allomorphy if, instead of having a unique
exponent, it has two or more contextually conditioned exponents. In this case, 1 ,
2 , ... are said to be allomorphs (of [F]).
Examples of allomorphy abound. In English, for instance, the nominal plural
has allomorphs (e)s and en, amongst several others:
(
en [N ox ]
(3)
[pl]
(e)s [N ]
This relation captures why the plurals of ox and fox are respectively, oxen and
foxes, not oxes and foxes or oxen and foxen: en is specified as the way the plural
is pronounced in the presence of ox, with (e)s used if no specific pronunciation
is otherwise specified. (The latter type of exponent is known as the default or
elsewhere item, capturing the fact that it is used where other, more contextually
specified items are not required.) Tables 13 present three further examples from
Australia, Austronesia, and Papua New Guinea, concerning genitive marking, accusative marking, and person in numberless pronouns, respectively.1
Superficially considered, allomorphy might seem a somewhat exceptional process, one that should be quite peripheral to the ambit of theories of grammar.
Consider the English plural (3) and the Imonda pronoun (table 3): in both, there
is a clear majority pattern (s-suffixation for English, f -suffixation without other
change for Imonda), with a second allomorph restricted to a small set. The same
holds to some extent for Hawaiian (table 2), as the majority of nouns are common
nouns and can even be true in cases, like Kalkatungu (table 1), where allomorphs
1
In all our examples, we have adhered, in so far as possible, to Carstairs (1987) useful heuristic
for identifying allomorphy, namely, that an irregular form occurs with otherwise regular inflection:
i.e., suppose that [F] and [G] ; then, if [FG] is pronounced , [G] can be said to have
allomorphs and ; if, however, [FG] is pronounced , then it cannot, without further argument, be
claimed that is an allomorph of [G] (or [F]) as it might be an exponent of the pair simultaneously.
For a dissenting view, see Trommer (1999); see also section 3.5.3 below.
Ua honi au ...
Ua honi au ia Lani
Ua honi au ia ia
Ua honi au i ka wahine/moa
I kissed ...
I kissed Lani
I kissed it
I kissed the woman/bird
Pronoun
ka
p l
ne
ehe
Emphatic/Reflexive
ka-f
ple-f a
be-f
ehe-f
1ex
1in
2
3
The first inclusive alternation p l ple is presumably phonological, as, despite presence of
word-final -lf (e.g., nffe-ual-f search-dl-pres, p. 114; e-ll-uagl-f dl-talk-do-pres, p. 108),
we cannot find any complex codas on monosyllables in schwa.
e
Seiler 1985: 44
a.
b.
c.
What is striking about these sentences is that the number marking is identical on
noun and verb: su for dual, sko for paucal, s for plural. However, it is surprising
that one should find this surprising: if agreement relations simply share information between parts of the clause, then one might naively expect Biak-like identity
to predominate. Though we are not aware of any typological studies addressing
the issue, our impression is that it does not. Instead occasional matching between
adjective/participle and noun though never with the verb seems to be the predominant pattern, exemplified here by the plural in Hebrew (5) and Georgian (6) and
the inverse (which functions, here, like a plural) in Kiowa (7):
(5)
(6)
(7)
[ph - g e- d ]-g
e- yy
[buffalo-inv inv-be ]-rel.inv inv-disappear.pf
The buffaloes that there were disappeared
c
cc
cc
If this is so, then allomorphy is a near ubiquitous property of any language with
agreement across nominal and verbal, or other similarity disparate, categories.
In this chapter our aim is not to propose a theory of allomorphy. Instead we
have two more modest aims. The first, broadly speaking empirical aim is to clarify
the criteria according to which a given alternation can be classified as allomorphic.
In slightly more concrete terms, this involves determining when an alternation
should be regarded as resulting from a representation like that in (2) as opposed
to other factors, such as phonology, morphology, or syntax.
The second, broadly speaking theoretical aim is to elucidate what each part
of (2) can stand for. One such issue was hinted at the outset, where we said that
the notion of feature set, represented by [F], was to be understood loosely, so as
not to exclude whole syntactic structures. In fact, questions surround every part
of (2). Besides the relatively obvioussuch as the content of [F], of , and of
)even the more innocuous notation devices disguise contentful
context (
questionswhether there are limits on the number of allomorphs (...), how and
when the choice ({) between different allomorphs is made, whether exponence
replaces [F] with or whether adds the latter on (). Needless to say, many
of these questions are mutually dependent particularly as concerns context: for
instance, if replaces [F], then this impacts on what context can contain; the
relative timing on which different exponence relation act can also affect context,
supply it with or depriving it of information. Full elucidation of these issues is too
great a task within current confines. Therefore, our aim will be to enumerate the
theoretical issues, to highlight some cores claims, and to lay out some data that
strikes us as germane to future attempts to construct the theory of allomorphy.
Two final notes before proceeding. First, our discussionwhich is couched in
realizational implementation of what Hockett (1954) terms an item-and-arrangement
(rather than item-and-process) approach to morphologyaims to clarify what we
think the empirical domain (section 2) and theoretical range (section 3) of an
investigation into allomorphy ought approximately to be. However, as will become immediately apparent from our first example below, our understanding of
the empirical domain of allomorphy is hardly theory-neutral. This is, of course,
unavoidable: to paraphrase Fodor 1981, data do not come with labels declaring
I am for allomorphy, I am for phonology, and so such classifications are up
to the researcher. Second, some authors use the term allomorphy to cover any
case of variation in a morphemes surface form (e.g., Bye 2008); others use the
term suppletion to refer to multiple underlying forms (cf, e.g., Spencer 1991);
and yet others use these terms differently again, restricting, e.g., suppletion to
non-affixal allomorphy (e.g., Embick 2009). Here, we eschew the term suppletion and use allomorphy to refer only to differences arising from the existence
of multiple underlying exponents.
2.
Empirical characterization
Consider again the English plural. In (3), the entry (e)s abbreviates the three
surface-distinct endings of the plurals of grace/graze, grate, and grey/grade, viz.
grac/graz[i-z], grate[s], and grey/grade[z]. The ending en of oxen is equally surface distinct. However, according to (3), these four surface variants comprise only
two distinct allomorphs: en and the others. In this section, we will be concerned
with the basis on which such decisions are made. Specifically, in the following
subsections, we consider how and why true allomorphy must be differentiated
from superficially similar variation that in fact arises from phonology (section
2.1), morphology (section 2.2), and syntax (section 2.3). We will furthermore
address some more conceptual issues, concerning the role of diachrony in establishing allomorphy (section 2.1) and the status of free variation (section 2.4).
2.1.
Phonology
Continuing with the English plural, the reason that we do not take (e)s to constitute separate allomorphsthat is, not to require listing as separate exponents
of [pl]is that application of regular phonology is sufficient to derive the surface variation from a single underlying form. More generally, we propose that
the notion of allomorphy applies only to underlying forms of exponents. That is,
if [F] has two surface-distinct exponents [1 ] and [2 ], then a necessary (though
not sufficient; see sections 2.22.4) condition for these to count as allomorphs is
the impossibility of deriving [1 ] and [2 ] from a single underlying // within the
phonology of the language.
That English plural [i-z][s][z] derive from one underlying exponent is probably familiar to most readers. Taking that form to be /z/, we posit -i-epenthesis
if the preceding segment is identical (modulo voicing and distributedness, e.g.,
grace/graze); if epenthesis does not apply, we posit devoicing if the preceding
segment is voiceless (grate); elsewhere, the underlying form surfaces (grey/grade).
That this is part of the phonology of the language is indicated, first, by the fact
that the phonotactic configurations thus avoided ([zs], [sz], [tz], [ds], ...) are absent
from codas throughout the language, whether derived or primitive, and, second,
6
grace
graze
grey
grade
grate
Plural
/Pres.
)
Past
)-[t]
-[i-z]
)
-[d]
)
-[z]
-[i-d]
-[s]
by the identical behavior of the homophonous present tense and, mutatis mutandis, past tense [i-d][t][d] (see table 4, noting that the above grace, etc., are
ambiguous between nouns and verbs).
Regarding the alternation [i-z][s][z] as non-allomorphic is not uncontroversial. Almost the whole of Matthews chapter on allomorphy (Matthews 1974, ch.
5; cf, Spencer 1991, ch. 4) is devoted to examples that we would not regard as
allomorphic precisely because regular operations of the phonology are capable
of yielding the surface variation from a unique underlying exponent, Matthews
prime example being the vowel-harmonic properties of Turkish number and case
suffixes. The basis of this position goes back to the Structuralist comparatively restricted notion of allophony. However, rather than enter into discussion of Structuralist phonology, consider how the two accounts fare with respect to the parallel
case in table 4. The anallomorphic approach posits underlying /z/ (plural/present)
and /d/ (past) and derives the surface variation, as already mentioned, by applying
[place1] and [+voiced]
phonological processes (informally, 7 -i / [place1]
7 [voiced] / [voiced]
) to the underlying forms. The allomorphic approach
posits as underlying forms the outputs of these phonological processes (viz., /i-z/,
/s/, /z/; /i-d/, /t/, /d/); and yet, it cannot do away with the information that the
phonological processes require, but repeats it as the allomorphic context (e.g.,
the context of [pl] s / [voiced]
is as in the devoicing process). Indeed,
whereas the anallomorphic approach represents this information once, as a process applying to all three underlying forms, the allomorphic approach must repeat
it triply, as the context for each stored allomorph. To assign such variation to the
lexicon, when phonology is perfectly equipped to handle it, seems a significant
7
loss of insight and, so, we restrict allomorphy to cases that cannot be derived by
regular phonology.
That said, it is not always trivial to decide where the boundary between phonological productivity and allomorphic listedness lies. First, there are cases where
the cumulative effect of individually attested phonological processes can lead to
surface alternations that appear to be non-phonological. For instance, Kiowa khi
thp exit exit.pf is derived by successive application of concatenation (khip) and the independently attested processes of vowel shortening, vowel lowering, tone simplification, and dental-velar switching (Watkins 1984, Harbour 2007:
123125). Much Caddoan morphophonology too yields this appearance (e.g.,
Melnar 2004). However, such problems generally relent on proper acquaintance
with the language in question.
A more difficult problem in distinguishing phonology from allomorphy arises
in connection to diachrony. A regular phonological process of a language may be
lost by descendant languages; however, some of the alternations induced by the
process may nonetheless remain. For instance, in Anglo-Saxon (to judge by careful perusal of Sweet 1896), all verbs with roots that end in short a and a single consonant have principle parts analogous to those of shake (viz., scacan to shake,
scc he shakes, scoc he shook, scocon they shook, scacen shaken). This
class is traditionally called the shake class and its pattern of vocalic alternations is distinct from all other classes (and also from purely suffixing so-called
weak verbs). Of the full list,2 a few maintain a similar pattern in Modern English:
the past tenses of shake and take are shook and took, though ache and bake no
longer yield ook and book. Furthermore, many new verbssuch as fake, flake,
snake, trache (i.e., tracheotomize)now fit the phonological frame of the modern
analogue of the shake class, whether as the result of sound change, neologism,
or borrowing. However, these, like many descendants of the shake class itself,
are simply treated as regularly affixing roots, without vowel alternation. The result is that what was once a well established phonological process of a language in
which vowel alternations were abundant has become an unpredictable alternation
that must be learned on a case-by-case basis.
Given the unpredictability of modern shake (in contrast to bake), one might
well be tempted to regard shake and shook as allomorphs, i.e., separate underlying
2
Viz., acan ache, bacan bake, calan be cold, dragan drag, go, faran go, gnagan
gnaw, hladan load, sacan quarrel, spanan instigate, tacan take, wacan be born; and
prefixed versions of these and other roots, such as a galan sound, a grafan engrave, forfaran
die, ofsacan deny charge, ongalan charm.
a.
b.
c.
d.
e.
The extremes of this continuum are uncontroversial: (8a) is the domain of phonology, (8e) is the domain of the lexicon and so of allomorphy.
Anglo-Saxon shake falls under (8c): systematic for phonologically like verbs,
but inapplicable to nouns like naman names, which undergo no analogous vowel
changes in declension, compounding, or category shift (witness invariant nam
in, respectively, namum names.dat, namena names.gen; namboc register,
namcuelce by name; namian mention by name, namnian call by name).
The remaining cases (8b, d) are handled in SPE by diacritic marking. For a
process with general application (8b), it is exempted elements that are diacritically
marked. For instance, obese is so marked (Chomsky and Halle 1968/1991: 174)
because obese obesity is excluded from the vowel laxing that affects serene
serenity and obscene obscenity. A similar phenomenon concerns lexical exceptions to general processes, such as unstressed vowel reduction in Catalan. There,
only [i], [u] and [ ] can occur in unstressed positions (9a); where affixation causes
stress shift, /a, , e/ neutralize to [ ] and /o, / to [u] (9b). However, some words
violate such constraints, which can be achieved by diacritic marking exempting
them from the relevant grammatical requirements and processes (9c).
e
e
e
e e
[p st 'na ]
carrot
[kunsuli 'sjo] consolidation
a.
(9)
We address in section 3.2 the use of readjustment rules and whether roots, as opposed to
affixes, are subject to allomorphy.
b.
stick stick.dim
weight heavy
blind blind.dim
put.3sg put.1pl
forgive.3sg forgiven
c.
['boston]
['klase]
Boston
class
e
e
e
e
e
e c
rankings (an approach that parallels use of minor rules, that is, phonological rules
that apply to a reduced set of items; (Lightner 1968)). Yet another approach within
OT appeals to lexically indexed constraints: exceptional items are indexed for the
application of specific faithfulness constraints (e.g., It and Mester 1999, 1999)
or of specific markedness constraints (for extreme cases where a specific phonological process seems to apply to a very small set of lexical items or specific morphemes; e.g., Pater 2000). Comparisons between these types of approaches can
be found, for instance, in Inkelas and Zoll (2007) or Pater (2009). What all share
is the idea that all the alternations discussed here are to be accounted for within
the phonology and without resorting to allomorphy. However, for cases like (8c),
especially where a single phonological process applies to a very reduced set or
even to just one morpheme, Mascar (2007) argues that allomorphy is indeed the
correct approach. In particular, he argues for this type of analysis for random
classes of lexical items, which do not constitute phonologically or morphosyntactically natural classes. However, he argues, even in cases where phonology does
not derive the surface variation from a single underlying form, it may nonetheless
have a role to play in resolving competition between the allomorphs.
We have seen that the situation in (8b), the case of a general phonological
process with a few exceptions, is generally handled through some diacritic marking on the exceptions that prevent them from undergoing the process. We have
also seen that the situation in (8c) is sometimes treated as phonologically conditioned allomorphy. However (8bc) are actually part of a continuum: one can
find, from a general process with few exceptions to an extremely limited process,
all sorts of phenomena that could be viewed as processes with more and more exceptions. The question arises, then, as to where to how to distinguish phenomena
that should be treated as phonological processes from ones that should involve
allomorphy (and, also, whether numbers of exceptions should be counted in terms
of types or tokens).
Of course, extensionally equivalent I-grammars could treat the same surface
variation as allomorphy for one individual and as phonology for another. However, an assumption implicit in most of the foregoing discussion is that I-grammars
minimize lexical storage and maximize rule-based coverage even for potentially
minor regularities. If so, then the scope for divergence between I-grammars may
be quite narrow. To justify this assumption, it is sufficient to test either the general
hypothesis (of list minimization, process maximization) or its specific relation to
the issue of minor phonological alternations. Of the wealth of psycholinguistic
literature on this topic, we pick just a few examples. First, neurolinguistic studies
(e.g., Stockall 2004) have argued that English morphophonological irregularity
11
is indeed synchronically decomposable and phonological. Second, psycholinguistic production experiments show that minority conjugations in English are generalized to cover nonce verbs, a fact easily explained if there is a rule to which
they can be made subject (though the interpretation of this behavior is subject
to debate; see, e.g., Bybee 2007). Third, acquisition of a low frequency vowelalternating verbs in English is apparently faster when the alternation has already
been mastered for another verb, again suggesting that the verbs are not acquired
as isolated exponents, but as instances of more general phonological processes
(Yang 1999).
Even if we determine, however, that shake and take are not subject to allomorphy, it must be recognized that, as the pool of irregular forms becomes smaller
and smaller, the likelihood that I-grammars will settle on allomorphy increases.
The precise point at which this occurs is far from obvious(8) is, after all, a
continuum, not a inventory of discrete typesand the debate is likely to turn on
psycholinguistic, and possibly neurolinguistic, evidence. Such evidence is likely
to be of interest to more mainstream linguistic concerns than pertain just to an
eventual theory of allomorphy. For instance, opaque phonological processes are
much more easily handled in serialist frameworks than parallelist ones. If the
psycho- and neurolinguistic research suggests that opacity triggers storage of allomorphs, then this may well favor theories in which opacity is achievable only
at high cost; if the finding is the reverse, then the opposite conclusion might be
upheld.
2.2.
Morphology
Just like phonology, the morphological system of a language can give rise to alternations that have the appearance of allomorphy. In this subsection, we exemplify
this and discuss criteria by which the two can be teased apart. The main example
we will focus on is the well known alternation between le and lo in Spanish.
The Spanish clitic system includes third person clitics for dative singular, le,
accusative, lo(s) and la(s), and reflexive, se. However, as Perlmutter (1971) observes, these do not cooccur as expected: a sentence like I gave him (le) it (lo),
which demands both a dative and an accusative clitic, in fact uses the reflexive and
accusative (se lo).
(10)
a.
Le
doy esto
dat (to him) I give this
I give this to him
12
b.
c.
d.
Lo
ve
acc (him) he sees
He sees him/it
Se
ve
refl (himself) he sees
He sees himself
Se
lo
doy
refl (to him) acc (her) he gives
I give it to him
To account for this non-reflexive use of se, Bonet (1991) proposes two features
structures for le and se according to which reflexive se realizes a proper subset of
dative le. Simplifying matters for the sake of illustration, we can use the following:
(11)
[cl dat] le
[cl] se
To force se to appear where, semantically, le is expected, as in (10d), Bonet proposes an impoverishment (feature deletion) rule, triggered when the dative clitic
precedes another clitic. On current terms:
(12)
dat 7 / [
] [cl ...]
The effect of (12) is to reduce the feature bundle to be realized to something big
enough only for se. On this analysis, the lese alternation is regarded as the effect
of a morphological operation.
However, for the data just given, an allomorphic account seems equally apt.
Specifically, let us suppose, roughly following Sportiche (1996), that clitics are
attracted to a particular clausal position, XP. Given that clitics precede finite verbs
(10) but follow imperatives (Dselo Give him it!), we can assume that XP intervenes between Force (the locus of imperatives) and T (the locus of finite verbs):4
(13)
a.
b.
This may be a simplification. If tenseless participles such as dando giving occur in functionally truncated clauses, the order dndo-se-lo giving him it suggests that the clitics are initially
below TP and move up to XP only when that head is present. Alternatively, the position of the
clitics in dndoselo might be taken to be the same as in finite clauses, with the participle moving
into the C domain.
13
We can then posit the following vocabulary item, in part recapitulating Perlmutters 1971 analysis:
(
le
X0
(14)
[cl dat]
se
On this account, the dative clitic le must be linearly adjacent to the head that
attracts the clitics. When another clitic, such as lo, intervenes, the allomorph se
emerges.5
Restricted to the Spanish data, it is hard to argue for the superiority of the
morphological approach over the allomorphic one. However, in crosslinguistic
perspective, the two can be teased apart. The contextual restriction on le is a
purely parochial fact about a single language: it provides no expectation or insight (beyond inheritance of historical accidental) as to how other languages might
work. Impoverishment, on the other hand, has been argued to be a reflex to excessive markedness (Noyer 1992, Nevins 2008) and, though different languages
may tolerate different thresholds of markedness, markedness is often consistent
crosslinguistically. Consequently, we are led to expect similar effects in other
languages.6
In fact, non-transparent realization of dative plus accusative clitics is a feature
of other languages too. Particularly interesting are the Italian dialects presented
by Manzini and Savoia (2005). They show that in exactly the configuration where
Spanish, or the dialect of Lconi, Sardinia, resort to reflexives, others, such as
the dialect of Nociglia in Apulia (15), use a partitive, and yet others, such as the
SantElia a Pianisi dialect of Molise (16), a locative:
5
14
(15)
Nociglia (Apulia)
(Manzini and Savoia 2005: 105)
a. li
'daj
'kwistu
dat (to him) he gives this
He gives this to him
b. la
'viSu
acc (her) I see
He gives her/it to him
c. n
li
'daj
'd i
part (of them) dat (to him) he gives two
He gives two of them to him
d. n
la
'daj
part (to him) acc (her) he gives
He gives her/it to him
(16)
e c
yc
15
a.
b.
Emhgyd
3sgD+3an.plO-know
He knows them (people)
Emgp
3sgA+3an.plO-hit.pf
He hit them (people)
cc
(17)
Syntax
Although perhaps less obvious, syntax too can induce alternations that resemble
allomorphy, a problem that is again remedied by paying closer attention to proper
syntactic analysis before positing allomorphic relations. In fact, even when allomorphy is to be posited, it can still be necessary to attend to syntactic behavior
7
The fact that the non-transparent realization is always another clitic/agreement form argues
against the allomorphy approach too, as originally noted by Bonet, contra Perlmutter.
16
in order not to misdescribe the context under which the allomorphs are used. We
illustrate these problems with examples from Yimas and Turkish, respectively.
In Yimas verbal morphology, there are a number of cases where arguments that
at other times cooccur with an overt morpheme on the verb fail to do so. Some
such zeroes have a clearly syntactic basis. For instance, there is an anti-agreement
effect on elements in various A-bar constructions.
(18)
a.
b.
(19)
a.
b.
The (b) examples show the agreement patterns typical of a given argument structure. The (a) examples show the same argument structure with the direct object
relativized. Observe that, where (b) has agreement for the object (na-, k-), the (a)
examples have none. Instead, they have only a that-like element (m-) prefixed to
the verb and an appropriate gender/number suffix (such discontinuous marking is
typical of Yimas; for an analysis, see Harbour 2008).
In contrast to syntactically determined zeroes, the language also has allomorphically determined zeroes. For instance, third person following the negative prefix ta in a (di)transitive verb is realized as either pu or as . The conditioning factor
is whether the following agreement morpheme begins with n (Harbour 2008: 201).
(20)
a.
b.
17
(21)
a.
b.
a.
b.
panmal [manpa
m- tu- t- ]
mani
crocodile j reli -kill-pf-sgi
the man who killed the crocodile
namat
[aykum m- tpul-c- um]
person.pl woman.pl j reli -kill- pf-sgi
the prople who hit the women
Given that this cannot be an anti-agreement effect (crocodile and women are
not in A-bar configurations), and given its similarity to negative-conditioned zeroes (there is a prefix, m/ta, in both cases; there is no subsequent n-initial agreement morpheme; zero occurs for what would be the left-most agreement morpheme of the unnegated verb), it is reasonable to consider this to be an allomorphic effect:
[tr]
[rel]
(23)
[3O]
.
..
However, Phillips (1993) argues that the zero here is syntactically determined:
presence of object agreement would, he argues, constitute a Relativized Minimality violation (given the morpheme order rel-3O-3A-V and assuming, with Foley,
that agreement morphemes are argumental). Interestingly, Phillips has supporting evidence for a syntactic approach. There is one case where anti-agreement is
suspended, namely, when the agent in the configuration 3O-3A-V is questioned
(Who hit him?). Phillips argues that independent constraints (similar to the EPP)
prevent omission of 3O here and so 3A is realized resumptively, again to prevent a
Relativized Minimality violation. Considerations of space prevent us from laying
out the argument in detail, however the pertinent point should be clear: even if
18
Free variation
Having raised the phenomenon of free variation, we wish briefly to point out that
this too might be regarded as a form of allomorphy, as suggested by Adgers work.
Examples of free variation are the Spanish imperfective subjunctive (table 5)
and the Catalan for nothing:
(24)
Juan quera
que cantaras/cantases una cancin
John want.3sg.impf that sing.2sg.impf.sjnct a song
John wanted you to sing a song
19
(25)
No vol
re/res ms
neg want.3sg nothing else
(S)he doesnt want anything else
One can regard this as the limiting case of allomorphy with regards to informativeness of context. Generally, the contextual specification is insufficient to determine
which allomorph should be used. In the case of the current examples, there does
not seem to be any contextual information, so either can be used (subject, doubtless, to extragrammatical, sociolinguistic factors):
(
ra
(26)
[impf.sjnct]
se
(
re
(27)
[nothing]
res
Instances of free variation appear rare, at least in comparison to contextual allomorphy. Languages frequently move from free variation to situations in which
each variant has a specialized context of occurrence. Carstairs (Carstairs 1987:
31) refers to this tendency in the domain of inflection as the Inflexional Parsimony Hypothesis, and he suggests it is relevant also for derivational morphology
(Carstairs 1988). However, it should be noted that Adgers analysis of Bucky English, which introduced this discussion, argues that free variation too can be an
intergenerationally stable state (cf, Ciarlo 2009).
3.
Theoretical issues
So far, we have only been concerned with the data that should be ascribed to
allomorphy. We have yet to say what a theory of that data consists of. Consider
again our initial schema, repeated below.
1 Context1
Context
2
2
(28)
[F]
..
...
.
As outlined in the introduction, every one of the notational devices in (28) raises
a set of contentful questions. These are far too numerous, and some of them far
too complex, to be treated in the current chapter. Immediately below, we lay out
some of the questions and in subsequent subsections, we examine in more detail
20
what strike us as some of the more important ones and giving indications about
core claims and current debates, as well as providing new data where possible.
3.1.
Questions
21
being one proposal within Minimalist work; Embick and Marantz 2008, Embick
20099 ).
The idea that [F] can include whole parts of phrase structure is implicit in Hale
and Keysers (1993) notion of l-syntax, according to which certain roots or affixes
correspond in some cases to small subtrees, and is explicit in versions of Minimalism that reject the idea that features are bundled together at terminal nodes (as
in Distributed Morphology and earlier related theories Halle and Marantz 1993),
but are, rather, individual projections along the clausal spine (as in the so-called
cartographic approach to clause structure; see Caha 2009 for detailed morphological analysis within this approach). The notion that roots are subject to exponence
is not, however, the same as claiming that they are subject to allomorphy. Indeed,
this has been explicitly rejected (e.g., Halle 1973, Marantz 1997; see section ??).
The content of is also a matter of interest. Clearly, can contain segmental
information. However, semi- and suprasegmental information can also be exponents. For instance, Banksira (2000) discusses at length the floating segment
exponents of Chaha; for example, the impersonal imperative involves a floating
segment that docks as a labial onto the rightmost non-coronal, as in si-r f w be
scared.impsn.imp, ki-f wc open.impsn.imp, f w-ic untie.impsn.imp (p. 207). (See
Akinlabi 1996 for treatment of many such examples.) And in Kiowa, discussed at
more length below, many grammatical properties, including transitivity and number agreement, are expressed tonally, as in hb enter.intr, hbe enter.tr,
and em 3sg.refl, m 3pl.refl. However, other values of are more controversial.
Zero exponence (see Trommer this volume) is one such case, as are diacritic features (such as conjugation or declension class) introduced by specific roots (e.g.,
Chomsky and Halle 1968/1991) and phonological junctures (e.g., Adger 2006b).
One might even imagine that could include syntactic structure (which would
then itself be subject to exponence, as a way to capture syntactic periphrasis) or
that it could be a phonological process, such as reduplication or truncation (thus
moving away from item-and-arrangement to item-and-process models).
There are also questions of interrelations between different exponents. Continuing with issues just raised, one can ask whether or not segmental exponents
can have all varieties of semi-segmental and non-segmental allomorphs. Equally,
.
one can ask whether there is a limit on the number (..) of allomorphs that a single
[F] can have, a position suggested by Carstairs (1987) (see section 3.3).
Implicit in the brace notation ({) is the notion that allomorphs of a given [F] are
e
It should be noted, however, that the phase is at times an extremely low upper bound, if, as
the cited works argue, the category heads, n, a, v, merged directly with roots, are phasal.
22
in competition: all exponents are potential realizations and, for each occurrence of
[F], the correct one must be chosen. This raises the question of how and when this
competition is resolved. With regard to how, two key concepts are that the subset
relation on contexts determines an ordering relation on the choice of allomorphs
(the Subset, or Pan.inis, Principle; e.g., Halle 1997) and that the bottom element
of that ordering, if lacking in context, is used where no other allomorph is licensed
(the Elsewhere Principle; e.g., Kiparsky 1973a).10 For instance, given the subset
relation between the contexts in (29), it follows that 2 will be used only if 1
cannot.
(
[H]
[G]
(29)
[F] 1
2 [G]
However, it should be observed that these principles do not induce a well-ordering.11
For instance, if Context1 is
[pl] and Context2 is
[2], then there is no
subset relation and so the competition has yet to be resolved. One might imagine
that general principles of markedness determine the ordering here (if, say, person
features are the more salient to competition resolution). However, we are not
aware of any thoroughgoing arguments for this position (and, indeed, if impoverishment is any indication, person and number are not in any universal markedness
relation; Noyer 1992, Adger and Harbour 2008). This suggests that the notion of
competition resolution is not yet properly understood or else may be somewhat
arbitrary. In this light, it is interesting to note that a good amount of recent phonological work (see section 3.4) has argued for removing the choice between some
allomorphs out of the specification of context and into the general workings of the
phonology.12
A further question related to timing of competition resolution concerns whether
exponence is determined simultaneously for all heads within a given spell-out domain or whether exponence sites are targeted sequentially. Answers to this question, and the ways in which such answers are implemented, entail differences in
the information that is available to other targets of exponence. For instance, if
10
See Caha (2007) on use of the Superset, rather than Subset, Principle, and hence the idea that
the bottom element of the ranking has, in fact, the largest, rather than the smallest, context of all
available allomorphs.
11
Adger (2006a) explores the idea that variability arises precisely where there is no intrinsic
well-ordering of exponents.
12
As section 3.4 makes clear, assigning allomorph selection to the phonology is not the same as
positing a single exponent and using the phonology to derive its surface variants, which we do not
regard as allomorphy at all (section 2.1).
23
exponence proceeds sequentially, then, on a top-down sequence, the highest target of exponence can only be non-phonologically conditioned, whereas a root-out
sequence predicts this for the lowest target.
In fact, answers to this question, like many of those already raised, are not
independent of positions taken with regard to other issues. A simple example
concerns interpretation of the exponence itself (): different quantities and varieties of information are available depending on whether exponence is an exchange
of morphosyntactic for phonological information (e.g., Trommer 1999, Bobaljik
2000) or an augmentation of the former by the latter (e.g., Anderson 1992). Moreover, both of the differences just mentioned and that of the preceding paragraph
impact on the information available to context.
In fact, context raises a wide variety of questions. First, all the questions
about variety of information and feature types that arise for [F] arise equally for
context. Additionally, though, there are questions that arise only in connection
to context. One, already implied above, is that of directionality, namely, which
information (morphosyntactic and/or phonological) is available above and below
a particular [F]. Another is that of distance, that is, whether a conditioning context
can see only adjacent information and, if so, whether linear or structural adjacency
is the relevant notion; or, indeed, whether different values of [F] (roots, features,
phrases) might not have access to different types of context.
As is evident, there is an abundance of questions that a theoretical account of
allomorphy might, or must, address, far more than can be addressed here. Below,
therefore, we have selected some that strike us as important in one of several ways:
either because they have been subject to much discussion, or because they ought
to be subject to more, or because we believe that we have data that is particularly
relevant and might move current debate forward.
3.2.
24
In this section, we review some data germane to these approaches and question
whether they are indeed sufficiently general to motivate the view that root variation
is fundamentally different from allomorphy. For a recent study of the topic, see
Siddiqi (2009).13
The simplest form of variation that roots exhibit is that of minor phonological
adjustment. Examples include /i/[] (creepcrep-t, leapleap-t, as opposed to
seepseep-ed), //[o] (sellsol-d, telltol-d, as opposed to wellwell-ed), and
/i/[] (ringrang, singsang, as opposed to dingding-ed). These cases require
only a change to one or two vocalic features and so would constitute phonologically plausible processes. Chomsky and Halle (1968/1991) termed these readjustment rules.
Such rules lead one to ask how much readjustment a readjustment rule can
undertake. For instance, though singsang may represent a single featural difference, bringbrough-t and thinkthough-t clearly do not. Within Chomsky
and Halles system, which has the underlying segments [ x], it might nonetheless be possible to derive these from /brin/ and /inx/, via appropriate readjustments. However, it is implausible that such a derivation could exist for gowen-t,
amwas, iswas.
However, Marantz (1993), inter alia, suggests that such massive variation is
restricted to roots that form part of the languages functional vocabulary, such
as auxiliaries and light verbs. Though this is certainly the case of English go
and be, it is not clear that all such variation can be so explained. Consider,
for instance, number-conditioned allomorphy in Kiowa: four (individual-level)
roots display a singulardual/plural split and seven (stage-level) roots display a
singular/dualplural split. Some examples are given in table 6 and, with the exception of small, no well constrained readjustment rule suggests itself. However,
it is also unclear that these roots can reasonably be termed functional. Some belong to semantic categories that are not infrequently suppletive for number (such
as predicates of bodily position and motion; Mithun 1988: 232), but it is doubtful
that this alone is sufficient to classify these predicates as functional. Certainly,
none is an auxiliary or light verb nor are they particularly well endowed with idiomatic meanings or senses. (This argument applies most forcefully in languages
where such predicates are relatively numerous; see in particular Hale, LaVerne,
and Pranka 1990 on Hopi and Tohono Oodham, f.k.a. Papago.)
However, Mithun (1988), concentrating precisely on North American languages, suggests that number-sensitive variation represents a real semantic dif13
(Add to main text: embick and noyer, oup: p 296, embick and halle.)
25
big
small
tall
short
sg
dl/pl
t
bn
sy n syn
kyy kn
x
xd
pl
k!l
p!l
th
z
child children
rebnok deti
rebnka detej
rebnku detjam
rebnkom detmi
rebnke detjax
man
men
c elovek ljudi
c eloveka ljudej
c eloveku ljudjam
c elovekom ljudmi
c eloveke ljudjax
ference between the verbs, one that we might translate into current terminology
as selection. On this view, the feature [singular] does not condition allomorphy
of t versus bn big, but t (or bn) is the realization of a version of the root big
that selects for a [+singular] (or [singular]) complement.
If we accept that all these defenses of root variability are legitimate, then, in
order to convince ourselves whether roots display allomorphy, we must discover
whether there are roots that are not plausibly functional, that undergo alternations
that are not mere phonology, and that cannot be said to select a complement with
the property to which they are allomorphically sensitive. Although this may seem
a tall order, such cases do apparently exist.
The simplest way to avoid the issue of complement selection is to leave verbs
and turn to nouns. Consider such variation as the Russian for child(ren) (table
7). The variation between the roots rebnok and det cannot be attributed to selection of a complement, as there simply is no complement, and it is clearly not the
result of minor readjustment. Furthermore, child does not plausibly seem to be a
26
functional, or light noun. That is, consider, in contrast, the behavior of Russian
man. As table 7 shows, man is just as variable as child. However, in a number of languages, man assumes a partly grammaticalized status. For instance, in
Scottish Gaelic, where there are two words for children, one count (pistean),
one mass (cloinne), person is used as a classifier when numerals cooccur with
the mass noun:
(30)
Scottish Gaelic
a. cig pistean deug
five child.pl teen
fifteen children
b. cig daoine cloinne
deug
five person.pl child.gen.sg teen
fifteen children
More subtly, in German, man may be used in the singular after a numeral. In
such cases, it is functionally reduced, not only in having no number marking, but
also in resisting adjectival modification; and it is semantically bleached, referring
not only to men, but also women and children (its usual plural, which may be
adjectivally modified, refers only to men).
(31)
German
(Domenica del Vecchio, p.c.)
a. Es sind fnfzehn (*unglckliche(r)) Mann ums leben gekommen
expl aux fifteen unfortunate.sg/pl man perished
Fifteen (*unfortunate) people (men, women, or children) perished
b. Es sind fnfzehn (unglckliche) Mnner ums leben gekommen
expl aux fifteen unfortunate.pl man.pl perished
Fifteen (unfortunate) men (not women or children) perished
Corbett (2007: 23) shows that Russian man displays a behavior somewhat similar to German (however, given the complexities of Russian numeral phrases, we
do not discuss the data here). It is, therefore, plausible that man is somewhat
grammaticalized as, hence, part of the functional vocabulary of some languages.
Consequently, its allomorphic behavior is possibly explicable in terms of grammatical allomorphy. For child, by contrast, we are unaware of any data that
could make such an analysis possible. (See Corbett 2007 for more such examples, including the extremely non-functional, but nonetheless allomorphic, pair
bicibodo sack corner(s), and for further discussion of morphologically regu-
27
saw
will see
faca
faic
h
[fa ca]
[f Ckj]
chunnaic ch
[xU
n Ckj] [Ci:]
played
will play
do chluich cluich
dep
[d xlUIC] [klUIC]
chluich
cluichidh
ind
[xlUIC]
[klUICi:]
e
3.3.
Number of allomorphs
To our knowledge, no proposals have been made that directly claim limits on
the number of allomorphs that may cooccur in a lexical entry. And languages
with high numbers of noun classese.g., the ten or so phonologically conditioned
allomorphs of the inverse number suffix in Kiowa (Harbour 2007: 55) or the
thirteen or so morphologically conditioned allomorphs of both singular and plural
28
in Arapesh (Fortune 1942: 48)might well lead one to question whether there are
empirical grounds for any such claims. However, some proposals by CarstairsMcCarthy do touch on this issue indirectly. As these issues have been discussed
more thoroughly elsewhere, we touch on these points on briefly, but with some
new data and, we hope, insight.
Carstairs 1987 and Carstairs-McCarthy 1994 propose, respectively, the Paradigm
Economy Principle and the No Blur Principle. The point of departure for both
principles is the observation that languages have the resources for far more noun
classes that they actually attest. For instance, for Latin nouns (Carstairs 1987:
66f), there are various nominative singular suffixes (domin-us, duc-s (written dux),
bell-um) and various nominative plural suffixes (domin-, duc-es, bell-a). If nominative singular formation and nominative plural formation are independent variables, then {-s, -um, -us} and {-a, -es, -} produce 3 3 = 9 logically possible
noun classes. Each set of case-number endings increases the number of logically
possible classes still further: ablative singular {-e, -o} increases it to 18; ablative
plural {-bus, -s}, to 36; genitive singular {-, -is}, to 72; ... Quite familiar languages have the resources for hundreds of noun (or verb) classes. The motivating
observation for the Paradigm Economy and No Blur Principles is that languages
have nowhere near this number.
This observation relates to the number of possible allomorphs as follows.
Carstairs argues that the number of noun/verb classes in a language can be limited
to about the right number if we can rule out paradigms such as the left-hand part
of table 9. According to Paradigm Economy (simplifying slightly), the number
of classes cannot exceed the number of allomorphs in the paradigm row with the
greatest number of distinctions. In (31), each row makes only two distinctions, so
there should only be two classes. Now, if correct, this principle is extremely difficult to render in our model. We have assumed that each exponence relation is an
autonomous entity. Here, however, one can only know that the exponence relation
below is ill-formed in virtue of the fact that there is no other number (dual, plural,
paucal, ...) in the language with three distinct allomorphs.
( Class 1
a
(32)
[F]
b Class 2/3
To incorporate such restrictions would therefore require some significant globalityor paradigm-based reconceptualization. It is therefore reassuring to observe that
Paradigm Economy is false. The Modern Hebrew nouns of table 9 (all of them
feminine) display precisely the distribution of affixes that Paradigm Economy
29
Class 2
b
c
Class 3
b
d
belly
egg
blessing
beten- beyc-ah brax-ah
btan-im beyc-im brax-ot
Class 2
a
Class 3
b
horse judge
nom pl hest-ar dmar-ar
gen sg hest-s dmar-a
Class 4
b
smith son
smi-ir syn-ir
smi-s son-ar
rules out.
Real though this counterexample is, one should not attach too much significance to the counterexemplification of Paradigm Economy, for two reasons.
First, Carstairs-McCarthy has moved from Paradigm Economy to No Blur, a
principle that rules out paradigms in which any row has more than one affix that
occurs in more than one cell. This rules out the following type of exponence
relation and the paradigm row in table 10.
(
a Context1
(Context2
(33)
[F]
Context3
b
Context4
This revised principle is not relevant to whether there is an upper bound on the
number of allomorphs, as it concerns, not numbers of allomorphs, but whether
multiple items may insert into multiple contexts. However, as this too falls well
within the purview of this chapter, it is worth observing, following Mller (2006:
30
Competition resolution
31
cho-lul
kim-ul
(*cho-ul) Cho-acc
(*kim-lul) Kim-acc
In Distributed Morphology, it is claimed that the choice of these two allomorphs must be determined at the point of Vocabulary insertion (see, for instance,
Embick 2009). The relevant Vocabulary Items would be as shown in (35).
(
lul V
(35)
[acc]
ul C
Reference to segmental information, and more crucially prosodic information
(needed for cases like the Tzeltal example below), is possible if one assumes that
phonological computation takes place inside out.
Within Optimality Theory, this type of approach is often rejected because of its
putative loss of generalization (to use Embicks 2009 phrase). That is, the context is stipulated: (35) just states that one allomorph is inserted after a vowel and
the other one, after a consonant. The Vocabulary Items do not express the fact that
the choice of allomorph gives rise to a more unmarked syllable structure (namely,
CV); that is, it is optimizing (in Pasters 2006 sense). If the choice of allomorph
is left to the phonology, phonological constraints that are needed independently
account for the distribution of the allomorphs. This version of Optimality Theory
accepts then the possibility of having multiple inputs. The tableau in (36), adapted
from Lapointe 1999, illustrates the choice of allomorph for the Korean accusative
suffix; other examples can be found in the literature. In the approach sketched in
(36), none of the candidates shown violates any faithfulness constraint because the
two allomorphs are present in the input; the choice is then left to markedness constraints, in this particular example the syllabic well-formedness constraints Onset
and NoCoda.14
14
32
(36)
R
R
Onset NoCoda
*
**!
Onset NoCoda
*!
*
*
At the other extreme, it is possible to find examples where the context for
allomorph selection is phonological but does not give rise to a more unmarked
configuration. A much cited example comes from Tzeltal. In this language, the
perfective suffix has two allomorphs: -oh after monosyllabic roots, and -h after
polysyllabic one (examples from Walsh Dickey 1999):
(37)
a.
b.
Contrary to the Korean case, it seems clear that the relation between the allomorph -oh (not -h) and monosyllabic stems in Tzeltal is idiosyncratic; it is this
allomorph, and not the other one, that selects a monosyllabic stem, and the choice
cannot be attributed to anything in the grammar or to any specific property of the
allomorphs. Therefore the distribution of the allomorphs must be explicitly stated
in either Distributed Morphology or a version of Optimality Theory that posits
multiple inputs.
The fact that not all instances of phonologically conditioned allomorphy can
be accounted for like the Korean example in (36)together with some other observations to be mentioned belowhas led some linguists to reject a phonologybased account to allomorphy for all cases, not just the Tzeltal type cases. As we
have seen, for Embick (2009), and more generally within Distributed Morphology, all competition has to be resolved at the point of Vocabulary insertion, as a
matter of principle. Paster (2006, in press) advocates a subcategorization model,
where all selection is sensitive only to underlying representations. In Korean, for
instance, the allomorph lul would subcategorize for a preceding vowel, and ul
33
In this particular case, the two allomorphs of each conjunction are inserted, but
with a precedence relation: [i]>[e]; [o]>[u]. Choosing the dispreferred allomorph
is penalized by the universal faithfulness constraint Priority. The tableaux below
illustrate how the combination of this faithfulness constraint and the markedness
constraint, OCP, account for the distribution of the allomorphs.15
(38)
OCP Priority
*!
*
(39)
OCP Priority
R
R
*!
In (38), choosing the allomorph [o] violates of OCP; to avoid this, the dispreferred
allomorph is selected. In (39), OCP violation is not an issue because the item
following the conjunction does not start with [o]; in such cases, the preferred
allomorph, the one that does not violate Priority, is chosen.16
An account with ordered allomorphs is also possible for the definite article
in Haitian Creole, a much cited case of apparently bizarre allomorph selection (a
perverse system, as Paster in press puts it). Again ignoring nasalization, the key
data is given below, with information about syllable structure:
(
a V
(40)
a. [the.sg]
la C
b.
For a more detailed analysis, which considers possibilities like diphthongisation or coalescence, see Bonet and Mascar (2006).
16
Again, positing ordered allomorphs is not incompatible with Distributed Morphology. The
Vocabulary Item for Spanish or, for instance, would be [or] {[o]>[u]} and the ordered pair
{[o]>[u]} be the input to the phonology.
35
can account for the distribution of the allomorphs (see Bonet, Lloret, and Mascar
2007 for a fully fledged analysis).
Wolf and McCarthy (2007) also assume a priority relation between allomorphs
but reject a parallel evaluation approach and the constraint Priority. Instead, they
argue that the preferred allomorph is tried first and that the candidate set includes
the null output, , which violates a single constraint, MParse (see Prince and
Smolensky 2004). If the null output is the optimal candidate the next allomorph is
tried. In the case of the Spanish conjunctions, for instance, for or plus otro, the
allomorph o is tried first. Since the candidate o otro violates OCP (ranked above
MParse), the null output is chosen. Given that the Lexicon contains another
allomorph, u, the other combination, u otro, can be tried out and, as this sequence
does not violate OCP, it beats the null output. In favor of their proposal, Wolf and
McCarthy (2007) mention cases like the comparative suffix -er in English, which
has a prosodic requirement: with long adjectives, like intelligent, the null output
beats the candidate with the suffix (*intelligenter); no other option is available,
given the lack of allomorphy for this suffix.17
An important question that arises with respect to phonologically conditioned
allomorphy is whether the phonological conditioning is based on the phonological representation of the context or on its surface form. When it is not based on
surface form, the conditioning becomes opaque. In many instances of phonologically conditioned allomorphy, there are no crucial differences between the two
representations and therefore this question does not arise. Few cases have been
discussed in the literature where the surface form, not the underlying form of the
context, is relevant. One example is provided by the personal article in Central
Catalan:18
(
l V
(41)
[the.masc.sg]
en C
In Catalan, initial sC clusters are not tolerated and schwa is inserted before the
cluster: [ ]scner scanner, [ ]snifar to sniff, [ ]spia spy. Epenthesis also
e
17
Wolf and McCarthy (2007) assume that periphrasis like more intelligent is obtained through
a different input syntactic structure, i.e., periphrasis and -er are not allomorphs.
18
In relation to the earlier discussion of diachrony, this example is an interesting one. Instead
of allomorphy arising as the residue of formerly active phonology (which once derived both forms
from a single underlying specification), the two definite article allomorphs here have their origins
in distinct words which have gradually semantically and morphosyntactically coalesced (distinguished now only by their phonological contexts of occurrence): en derives from Latin domine
master, l from the Latin deictic ille.
36
loc.sg
lis[c]-e
lis[c]-u
affects sC-initial proper names, but the allomorph chosen for the personal article
is l [l], not en [ n], as illustrated in (42).
e
a. l[ ]Smolensky
b. *en ([ ])Smolensky
e
(42)
In Distributed Morphology, this can be accounted for easily if one assumes that
Vocabulary insertion as well as the phonology apply cyclically. In order to obtain
l[ ]Smolensky, the proper name should be introduced first (or should be there
from the start, as assumed by Embick 2009), and, crucially, epenthesis should
take place before the personal article is inserted. In parallel OT multiple input
approaches, the solution is also straightforward because the vocalic context is
surface true.
Other cases have been claimed in the literature to be sensitive to the underlying
representation of the context, not its surface representation, because a phonological process has rendered the context opaque. In these cases the conditioning is
not surface true, and therefore poses problems to parallel models of OT. Several
instances of opaque allomorphy in different languages have been reported. Much
cited examples are provided by Polish (ubowicz 2007), Spanish (Lang 1990,
Aranovich and Orgun 2006), and Turkish (Aranovich, Inkelas, and Orgun 2005;
see also Paster in press). Let us illustrate the phenomenon with a Polish example
from ubowicz.
In Polish the locative singular suffix has the allomorphs [e] and [u]. Even
though [u] appears after prepalatal consonants, and [e] after nonprepalatals, the
distinction is rendered opaque because of a process of palatalization; on the surface the segments that motivated the choice are identical. Table 11 illustrates this
with two examples that constitute a near minimal pair. Obviously, this case cannot
be accounted for in the Standard parallel model of Optimality theory. ubowicz
(2007), however, argues for a parallel model within an approach that includes
constraints that penalize pairs of inputs that map onto the same output.
e
37
For cases of allomorph selection where only the phonology determines which one
of the allomorphs is chosen, he assumes, as is done in other work within Optimality Theory, that markedness constraints are the ones forcing the selection of
each allomorph. For cases where one of the allomorphs is the preferred one, the
other one being selected only when there is some phonological conflict (cases
of arbitrary preference), he assumes, and must assume, that the two allomorphs
do not express identical morphosyntactic information. For the ergative suffix in
Dyirbal, for instance, he claims that one of the allomorphs, ku, expresses two
19
For most of his thesis, Wolf assumes that morpheme realization proceeds from the root outwards, but he entertains the idea that it could proceed outwards or inwards, based on some evidence
from Italian and Southern Zaria Fulfulde. See section 3.5.1.
38
Feature type
phonological
phonological
morphosyntactic
morphosyntactic
phonological
phonological
morphosyntactic
morphosyntactic
Locality
adjacent
long distance
adjacent
long distance
adjacent
long distance
adjacent
long distance
Example
Georgian (44)
none?(see p. 41)
Latin (p. 46)
Kiowa? (50)
none?(see note 21)
none?(see note 21)
Georgian (45)
Itelmen? (47)
different sets of features, [free], and other features related to ergative ([oblique
+structural +superior]), while the other allomorph, ku, does not express [free].
Although this account may be feasible for Dyirbal, it is very difficult to imagine what morphosyntactic features may distinguish allomorphs in other cases of
arbitrary preference, like the Spanish conjunctions y/e and and o/u or.
3.5.
Context
The discussion so far has shown that allomorphy can be both morphosyntactically
and phonologically conditioned. Moreover, the sensitivities have been both inward (i.e., a higher affix dependent on more deeply embedded information; table
1) and outwards (i.e., a more deeply embedded affix dependent on structurally
higher information; table 3). And, in all cases, nothing has intervened between
the allomorph and the information on which its choice depends. It is natural to
ask which of these conditions are necessary for allomorphic dependencies. That
is, one can imagine a set of independent variables which divide up the space of
logically possible allomorphic dependencies: inwards versus outwards sensitivity; local versus long distance sensitivity; phonological versus morphosyntactic
sensitivity. These possibilities are laid out in table 12 with examples where we
believe these can be provided.
As we discuss in passing below (section 3.5.3), one might further refine these
questions by asking whether all values for [F] permit the same options (for instance, one can imagine that roots are subject to more local conditioning than
affixes). Similarly, one might ask whether the notions of locality are the same for
39
The main focus for our discussion, both here and below, will be Bobaljiks claim
that the varieties of attested sensitivities can be derived from the hypotheses that
exponence is not an all-at-once operation, targeting all exponence sites within a
domain simultaneously, but proceeds from one target to the next starting at the
root and cycling outwards, and that exponence is replacive, that is, if [F] ,
then, once is present, [F] no longer is (cf, Trommer 1999). That is, if we have
three targets of exponence, [K [L [M ...]]] and [M] , then, when [L] is target
for exponence, the structure will be [K [L [ ...]]], entailing that [L] has access to
the (but not the morphosyntactic information of which it is the exponent) and
[K] (but not the phonological information of its eventual exponent). In its simplest
form, this predicts that all and only inwards sensitivities will be phonological and,
conversely, that all and only outwards sensitivities will be morphosyntactic.20, 21
Georgian datives provide a simple illustration of inwards phonological and
20
More complex patterns of sensitivities can be derived given other assumptions. For instance,
if exponence targets terminal nodes, and if one assumes an X-bar-like syntax, then the morphosyntactic information in category labels (MP in the main text example) remains visible even when exponence has replaced the terminal node itself. Thus, one could have allomorphy tense for aspect,
but not for a particular value of aspect, such as perfective or imperfective.
21
As with the definition of allomorphy itself, one should not be confused into thinking that
outwards sensitivity to phonological information is completely impossible: such variation as is
derivable by regular phonology is, of course, permitted. For instance, the quality of the root vowel
in the declension of Icelandic fjr fjord (Einarsson 1945: 36) is clearly conditioned by the
suffixal vowel (not by such heterogeneous sets as {dat.sg, nom.pl, acc.pl}) and can be handled,
along with deletion of the glide, by slight, properly circumscribed phonological alternations.
nom
acc
dat
gen
sg
fjrur
fjr
firi
fjarar
40
pl
firir
firi
fjrum
fjara
a.
b.
c.
a.
b.
If Bobaljik is correct that outward-sensitive phonological conditioning is impossible, then, in particular, long-distance outwards-sensitive phonological conditioning will be unattested, just as we have (tentatively) found long-distance inwardssensitive phonological conditioning to be. Therefore, let us turn to long-distance
morphosyntactic conditioning. Here we consider two examples, that operate in
different directions, one Bobaljiks own, from Itelmen, the other, from Kiowa.
Bobaljiks example concerns phi-feature sensitive. He argues that three parts
of the Itelmen verb are sensitive to phi-features of two potential sources of agreement, subject and object, and that these three heads, label A, B, C are hierarchically structured with suffixal C innermost and prefixal A outermost:
22
For this reason, it is hard to regard as relevant such foot-sensitive allomorphy as the
Kalkatungu operative, i.e., ergative, where Blake (1969: 33) describes the conditions on u
t versus
()ku as requiring, inter alia, more than two versus exactly two syllables. However, at the level
of feet, one can distinguish these as contrasting non-word-initial versus word-initial feet. Thus
ignoring segments, this seems to involve a simple adjacency sensitivity at the level of feet. Like
comments apply to the Tzetzal data of (37).
42
(46)
[A [[verb C] B]]
In a transitive verb, the subject and object phi-features to which these positions
are sensitive are, for [A], subject only; for [B], primarily those of the object and,
potentially, a subpart of the subjects too; and for [C], both simultaneously, depending on the phi-features in question. Thus, in a maximally rich verb, [A] will
reflect properties of the subject, and [B] and [C], of subject and object, as shown
below (examples from Volodin are via Jonathan Bobaljik, p.c.; 0S stands for
impersonal subject):
b.
c.
a.
(47)
43
A typical example of their use is shown below, where, for convenience, we have
chosen a pair on which transitivity is overtly marked:
cc
cc
b.
hb- e- t
/*t!
enter-tr-mod(tr)/*mod(intr)
will bring in
hb- - t!
/*t
enter-intr-mod(intr)/*mod(tr)
will come in
cc
a.
cc
(49)
cc
cc cc
b.
/*t!
hb- e- guu m -t
enter-tr-distr-neg- mod(tr)/*mod(intr)
will not bring in at different times/locations
hb- - guu m -t!
/*t
enter-intr-distr-neg- mod(intr)/*mod(tr)
will not come in at different times/locations
cc cc
a.
cc
(50)
Here, the modal suffix continues to show allomorphic sensitivity across two intervening suffixes. Interestingly, in the absence of the distributive, the form of
the negative changes and the expression of transitivity is absent from the surface
string. Consequently, (in)transitivity is registered only by the long-distance inwards sensitivity of the modal:
cc
cc
cc
cc
b.
/*t!
hb- - t
enter-neg-mod(tr)/*mod(intr)
will not bring in
/*t
hb- - t!
enter-neg-mod(intr)/*mod(tr)
will not come in
cc
a.
cc
(51)
Interesting though these Itelmen and Kiowa examples are, they are not unproblematic. Bobaljiks theory rules the latter type out, as, the relevant transitivity
features should be absent by the time the modal is targeted for exponence. One
44
might, as a result, reject just the portion of his proposals that views exponence
as replacive. Alternatively, one might maintain this, but claim that the relevant
information occurs both low (where it is expressed) and high (where it conditions
allomorphy). Such duplication of information is, of course, common in syntax: it
is the stuff of agreement relations. So, one can maintain a replacive view of exponence if one is willing to posit an Agree relation between transitivity and modality, and between aspect and evidentiality. This move comes at a price, however,
in that there seems to be little grounds, either within Kiowa, or crosslinguistically,
for positing the necessary Agree relations.
At the same time, Bobaljiks own example is not immune from reanalysis.
Note that [C] comes directly after tense. If subjects in Itelmen move to, or are
in an Agree relation, with T, then [C] might be taken to be the pronunciation of
(some of) those features. Alternatively, if, contra Bobaljik, [B] is in fact a locus of
subject agreement features, just as [A] is, then this would make [C]s sensitivity to
subject features a simple matter of allomorphy for the structurally adjacent head
[B]. Supporting this view, Susi Wurmbrand has observed (J. Bobaljik p.c.) that,
under nearly all circumstances, the subject features expressed at [C] are a subset
of those expressed at [B].
As with all the phenomena discussed in this section, more examples must be
sought before firm conclusions can be drawn. Methodologically, we note that
the most robust evidence is likely to come from clitics or distinct heads hosting
agreement with arguments (as in Itelmen). Using such data precludes the concern,
raised regarding Kiowa, that there might be an ad hoc Agree relation between the
heads in question: this cannot arise for, say, subject and object agreement, because
they are the result of two such relations themselves and, if there were a further
such relation between them, their feature content would be identical.
3.5.3.
Adjacency-dependent allomorphy
45
phy are furthermore interesting in their own right. First, both root and affixal
allomorphs can be adjacency-dependent. Although we hesitate to suggest generalizations prior to proper typological sampling, we find it interesting that our
few instances of long-distance allomorphy involve only affixes and that, complementarily, all our examples of root allomorphy appear to be adjacency-dependent.
Second, given that exponence is the heart of the mapping from syntax to phonology, it is interesting to ask whether it is the syntactic notion of structural adjacency
or the phonological notion of linear adjacency that is relevant to allomorphic contexts. Again, we suffer from a paucity of examples that militates against general
conclusions; however, in at least one potentially relevant case (brought to our attention by Pavel Caha), it is clear that structural adjacency is crucial. That said, as
made clear below, we believe that the current state of morphological theory may
militate against any firm conclusions being drawn here.
We begin with an illustration of adjacency-dependent affixal allomorphy (which,
at the same time, fills another cell from table 12: inwards-sensitive, adjacencydependent, morphosyntactically conditioned allomorphy). The case in question
comes from Latin and concerns the expression of second person singular verbal
agreement (Adger, Bjar, and Harbour 2003). In the present perfect indicative
you have loved, this takes the form ist, as in ama-u-ist (love-pf-2sg). However, if any other exponent occurs between pf and 2sg, as in the perfect subjunctives ama-u-eri-s and ama-u-isse-s, then 2sg has the same (default) exponent as in
such non-perfect forms as the present and imperfect future indicatives ama-s and
ama-bi-s. These examples therefore show that ist is an exponent of 2sg which
is inwards-sensitive to a morphosyntactic conditioner, pf, but which crucially requires adjacency with that conditioner.
A similar example for roots is provided by Spanish people:
(52)
a.
b.
Of particular interest are such forms as populate versus popularize. Both are
verbal and, when the verbal affix is adjacent to the root, the latter takes the form
pobl, which it retains when further suffixes are added (52a). By contrast, if an
adjectival affix is adjacent to the root, the latter takes the form popul, which it
again retains under further suffixation (52b). Thus, when an adjectival affix disrupts adjacency between the root and a verbal affix, the latter cannot condition the
allomorph pobl (hence, popul-ar-iz-ar, *pobl-ar-iz-ar).
46
Interesting though the Latin and Spanish examples may be, they do not reveal which variety of adjacency, structural or linear, the allomorphy requires:
given that all the affixes are suffixal, structural adjacency obtains if and only if
linear adjacency does. The nearest we can find to an example that teases these
two conditions apart comes from Korean. In a thorough treatment of negation
in that language, Chung (2007) pays particular attention to the verbs know and
exist, which coalesce with negation (a process which, e.g., Trommer 1999 argues is a form of allomorphy). For instance, al-ass-ta (know-past-decl) negates as
moll-ass-ta (neg.know-past-decl), rather than as *an/ani/mos al-ass-ta (neg knowpast-decl), which would be expected given other verbs in the language (p. 115).
Interestingly, though, the negative of causativized know does not use the negative allomorph. Chung argues (p. 132) that the functional hierarchy of such verbs
is T > neg > cause > V (cf, Cinque 1999). This case is, therefore, analogous to
Latin: just as the special allomorph of 2sg is blocked when not adjacent to pf,
so, here, the special (coalesced or allomorphic) form of know is blocked when
not adjacent to neg. Korean differs from Latin and Spanish, however, in that neg
and cause are both linearly adjacent to V ([[neg [know cause]] pres]; p. 132),
and shows therefore, that linear adjacency, in the absence of structural adjacency,
does not suffice to produce the special negative forms: an/ani/mos al-li-ess-ta (neg
know-caus-past-decl), but *mol(u)liessta (neg.know-caus-past-decl) (p. 120).
Though this constitutes allomorphy only relative to a certain set of assumptions, the example is nonetheless instructive in showing how one can distinguish
structural from linear adjacency as a condition on allomorphy. Naturally, it is
possible that morphosyntactically conditioned allomorphy might be sensitive to
structural, and phonological conditioned allomorphy, to linear adjacency. However, if Marantz (1984; see also Embick and Noyer 2001) is correct that morphological processes, such as Merger, can establish adjacency relations absent from
the syntax, then it is likely to prove extremely difficult to distinguish linear from
structural adjacency: any circumstance under which linear adjacency holds might
be made into one in which structural adjacency does too, provided the requisite
morphological operations may apply.
To make this concrete, consider ordinal allomorph in English and Italian (data
from Michele del Vecchio, p.c.). In both, higher ordinals are derived by regular
suffixation to the basic numeral, for instance, venti/ventesimo and twenty/twentieth,
but some lower numerals form their ordinals irregularly, as in due/secondo (*duesimo) and two/second (*twoth). Moreover, the two languages are string-identical
for the number ventidue, twenty-two. However, they diverge with respect to the
corresponding ordinal, with Italian using the otherwise ungrammatical twoth,
47
Conclusion
In this chapter, we have sought to clarify the empirical grounds on which variation
form can be attributed to allomorphy, that is, to the existence of a choice of exponents for a given syntactic structure, as opposed to the operations of phonology,
morphology, or syntax. Furthermore, we have adumbrated the many theoretical
issues that a full account of allomorphy must address, emphasizing in particular
parallels between root and affixal allomorphy, the apparent lack of restrictions on
numbers of allomorphs, the role that phonology may play in determining how
competition between allomorphs is resolved, and the complex, interrelated set of
questionsdistance, directionality, adjacency, and feature typesraised by the
notion of context. As emphasized at the outset, our aim has not been to develop a
theory of all of these factors, but to clarify the facts and factors that such a theory
should address. If we have been successful, then hopefully our own observations
will be surpassed by such a theory in the near future.
References
Adger, David. 2006a. Combinatorial variability. Journal of Linguistics
42:503530.
Adger, David. 2006b. Post-syntactic movement and the Old Irish verb. Natural Language and Linguistic Theory 24:605654.
Adger, David. 2008. A minimalist theory of feature structure. Ms, lingbuzz.
Adger, David, Bjar, Susana, and Harbour, Daniel. 2003. Directionality of
allomorphy: A reply to Carstairs-McCarthy. Transactions of the Philological Society 101:109115.
48
Adger, David and Harbour, Daniel. 2007. Syntax and syncretisms of the
Person Case Constraint. Syntax 10:237.
Adger, David and Harbour, Daniel. 2008. Whi phi? In D. Harbour,
D. Adger, and S. Bjar, eds., Phi Theory: Phi-Features Across Domains
and Interfaces, 138, Oxford: Oxford University Press.
Adger, David, Harbour, Daniel, and Watkins, Laurel J. 2009. Mirrors and
Microparameters: Phrase Structure Beyond Free Word Order. Cambridge: Cambridge University Press.
Adger, David and Smith, Jennifer. 2005. Variation and the Minimalist Program. In Leonie Cornips and Karen Corrigan, eds., Syntax and Variation: Reconciling the Biological and the Social, 149178, Amsterdam:
John Benjamins.
Akinlabi, Akinbiyi. 1996. Featural affixation. Journal of Linguistics
32:239289.
Anagnostopoulou, Elena. 2003. The Syntax of Ditransitives: Evidence from
Clitics. Berlin: Mouton de Gruyter.
Anderson, Stephen. 1992. A-morphous Morphology. Cambridge: Cambridge University Press.
Aranovich, Ral, Inkelas, Sharon, and Orgun, Orhan. 2005. Opacity in
phonologically conditioned suppletion, paper presented at the Thirteenth Manchester Phonology Meeting.
Aranovich, Ral and Orgun, Orhan. 2006. Opacity in -ez/-eza suffixation.
In Timothy L. Face and Carol A. Klee, eds., Selected Proceedings of
the 8th Hispanic Linguistics Symposium, 116122, Somerville, MA:
Cascadilla Proceedings Project.
Banksira, Degif Petros. 2000. Sound Mutations: The Morphophonology of
Chaha. Philadelphia/Amsterdam: John Benjamins.
Blake, Barry J. 1969. The Kalkatungu Language: A Brief Description. Canberra: Australian Insitute for Aboriginal Studies.
Bobaljik, Jonathan. 2000. The ins and outs of contextual allomorphy, ms.
McGill University, Montral.
Bonet, Eullia. 1991. Morphology after syntax: Pronominal clitics in Romance. Ph.D. thesis, MIT, Cambridge MA.
49
Bonet, Eullia. 1995. Feature structure of Romance clitics. Natural Language and Linguistic Theory 13:607647.
Bonet, Eullia, Lloret, Maria-Rosa, and Mascar, Joan. 2007. Lexical
specifications and ordering of allomorphs: Two case studies. Lingua
117:903927.
Bonet, Eullia and Mascar, Joan. 2006. U u o e y o e. In Antonio Fbregas,
Natalia Curto, and Jos Mara Lahoz, eds., Cuadernos de Lingstica
XIII, 19, Madrid: Instituto Universitario de Investigacin Ortega y
Gasset.
Bybee, Joan. 2007. Frequency of Use and the Organization of the Lexicon.
Oxford: Oxford University Press.
Bye, Patrick. 2008. Allomorphy: Selection, not optimization. In Sylvia
Blaho, Patrick Bye, and Martin Krmer, eds., Freedom of Analsyis?,
6392, Berlin: Mouton de Gruyter.
Caha, Pavel. 2007. The shape of paradigms, GLOW XXX.
Caha, Pavel. 2009. The nanosyntax of case. Ph.D. thesis, Universtitetet i
Troms.
Carstairs, Andrew. 1987. Allomorphy in Inflexion. London: Croom Helm.
Carstairs, Andrew. 1988. Some implications of phonologically conditioned
suppletion. In Geert Booij and Jaap van Marle, eds., The Yearbook of
Morphology 1988, 6794, London: Routledge.
Carstairs-McCarthy, Andrew. 1994. Inflectional classes, gender, and the
Principle of Contrast. Language 70:737788.
Carstairs-McCarthy, Andrew. 2001. Grammatically conditioned allomorphy, paradigmatic structure, and the ancestry constraint. Transactions
of the Philological Society 99:223245.
Carstairs-McCarthy, Andrew. 2003. Directionality and locality in allomorphy: a response to Adger, Bjar & harbour. Transactions of the Philological Society 101:117124.
Chomsky, Noam and Halle, Morris. 1968/1991. The Sound Pattern of English. Cambridge: MIT Press, paperback edition.
Chung, Inkie. 2007. Suppletive negation in Korean and Distributed Morphology. Lingua 117:95148.
50
52
Inkelas, Sharon, Orgun, Orhan, and Zoll, Cheryl. 1997. The implications of
lexical exceptions for the nature of the grammar. In John McCarthy, ed.,
Optimality Theory in Phonology: A Reader, 542551, Oxford: Blackwell.
Inkelas, Sharon and Zoll, Cheryl. 2007. Is grammar dependence real? A
comparison between cophonological and indexed constraint approaches
to morphologically conditioned phonology. Linguistics 45:133171.
It, Junko and Mester, Armin. 1999. The phonological lexicon. In Natsuko
Tsujimura, ed., The Handbook of Japanese Linguistics, 62100, Oxford: Blackwell.
Kayne, Richard S. 1975. French syntax: the transformational cycle. Cambridge, MA: MIT Press.
Kiparsky, Paul. 1973a. Elsewhere in phonology. In Stephen Anderson
and Paul Kiparsky, eds., A Festschrift for Morris Halle, 93106, New
York: Holt, Rinehart, and Winston, Inc.
Kiparsky, Paul. 1973b. Phonological representations. In Osamu Fujimura,
ed., Three Dimensions of Linguistic Theory, 1135, Tokyo: TEC.
Lang, Mervyn. 1990. Spanish Word Formation. London and New York:
Routledge.
Lapointe, Steven. 1999. G. 1999. stem selection and OT. In Geet Booij and
Jaap van Marle, eds., Yearbook of Morphology 1999, 263297, Dordrecht: Kluwer.
Lightner, Teodor. 1968. On the use of minor rules in Russian phonology.
Journal of Linguistics 4:6972.
ubowicz, Anna. 2007. Paradigmatic contrast in Polish. Journal of Slavic
Linguistics 15:229262.
Manzini, Maria Rita and Savoia, Leonardo Maria. 2005. I dialetti italiani e
romanci: Morfosintassi generativa. Alessandria: Edizioni dellOrso, 3
volumes.
Marantz, Alec. 1984. On the Nature of Grammatical Relations. Cambridge,
Mass.: MIT Press.
Marantz, Alec. 1993. Implications of asymmetries in double object constructions. In Sam Mchombo, ed., Theoretical Aspects of Bantu Grammar, 113150, Stanford, CA: CSLI Publications.
53
Marantz, Alec. 1997. No escape from syntax: dont try morphological analysis in the privacy of your own lexicon. In A. Dimitriadis, L. Siegel,
C. Surek-Clark, and A. Williams, eds., PWPL 4.2, Proceedings of
the 21st Annual Penn Linguistics Colloquium, 201225, Pennsylvania:
University of Pennsylvania Working Papers in Linguistics.
Mascar, Joan. 1976. Catalan phonology and the phonological cycle. Ph.D.
thesis, Massachusetts Institute of Technology, Cambridge, MA, distributed by Indiana University Linguistics Club.
Mascar, Joan. 2007. External allomorphy and lexical representation. Linguistic Inquiry 38:715735.
Matthews, Peter. 1974. Morphology: An introduction to the Theory of
Word-Structure. Cambridge: Cambridge University Press.
McCarthy, John. 2007. Hidden Generalizations: Phonological Opacity in
Optimality Theory. London: Equinox.
Melnar, Lynette. 2004. Caddo Verb Morphology. Lincoln: University of
Nebraska Press.
Mithun, Marianne. 1988. Lexical categories and the evolution of number
marking. In Michael Hammond and Michael Noonan, eds., Theoretical
Morphology: Approaches in Modern Linguistics, 211234, San Diego:
Academic Press, Inc.
Mller, Gereon. 2006. Notes on paradigm economy. In Gereon Mller
and Jochen Trommer, eds., Subanalysis of Argument Encoding in Distributed Morphology, number 84 in Linguistische Arbeitsberichte, 161
195, Leipzig: Universitt Leipzig.
Nevins, Andrew. 2007. The representation of third person and its consequences for Person-Case effects. Natural Language and Linguistic Theory 25:273313.
Nevins, Andrew. 2008. Cross-modular parallels in the study of phon and
phi. In Daniel Harbour, David Adger, and Susana Bjar, eds., Phi Theory: Phi-Features across Interfaces and Modules, 329367, Oxford:
Oxford University Press.
Noyer, Rolf. 1992. Features, Positions and Affixes in Autonomous Morphological Structure. Cambridge, MA: MITWPL.
54
55
56
Contents
1
Introduction
Empirical characterization
2.1 Phonology . . . . . . .
2.2 Morphology . . . . . .
2.3 Syntax . . . . . . . . .
2.4 Free variation . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6
. 6
. 12
. 16
. 19
Theoretical issues
3.1 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Roots and allomorphy . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Number of allomorphs . . . . . . . . . . . . . . . . . . . . . . .
3.4 Competition resolution . . . . . . . . . . . . . . . . . . . . . . .
3.5 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.1 Feature types and directionality . . . . . . . . . . . . . .
3.5.2 Distance, directionality and morphosyntactic conditioning
3.5.3 Adjacency-dependent allomorphy . . . . . . . . . . . . .
20
21
24
28
31
39
40
42
45
Conclusion
48
57